GPT-5.6 Sol costs $5.00 per million input tokens and $30.00 per million output. Terra is $2.00 in and $12.00 out. Luna, the cheap one, is $0.20 and $1.20. Those are standard-tier prices for prompts up to 272K input tokens — go past that and input doubles while output rises 1.5x, which most pricing write-ups never mention. Two of the three moved on July 30, 2026.
Last verified: 11 August 2026.
I re-check this page monthly against OpenAI's own pricing page and changelog, not against other people's summaries. Per-token pricing is the fastest-rotting number in this industry and a stale table here costs you real money.
The GPT-5.6 pricing ladder
Everything below is the Standard service tier, in dollars per 1M tokens, from developers.openai.com/api/docs/pricing.
One thing you need to know before reading any GPT-5.6 price anywhere: OpenAI's table is split into two context bands. The header tooltips define them precisely — short context is ≤272K input tokens, long context is >272K. Almost every summary you'll find quotes only the short-context half. Here are both.
Short context (≤272K input tokens):
| Model | Input | Cached input | Output |
|---|---|---|---|
| gpt-5.6-sol | $5.00 | $0.50 | $30.00 |
| gpt-5.6-terra | $2.00 | $0.20 | $12.00 |
| gpt-5.6-luna | $0.20 | $0.02 | $1.20 |
Long context (>272K input tokens):
| Model | Input | Cached input | Output |
|---|---|---|---|
| gpt-5.6-sol | $10.00 | $1.00 | $45.00 |
| gpt-5.6-terra | $4.00 | $0.40 | $18.00 |
| gpt-5.6-luna | $0.40 | $0.04 | $1.80 |
The jump is not a rounding detail: input doubles and output goes up 1.5x the moment a request crosses 272K input tokens. If your workload is long documents — codebases, contracts, transcript piles — the long-context column is your real price, and budgeting off the short-context one understates your bill by around 2x on input.
There's a fourth column on OpenAI's table that most write-ups drop: cache writes, which cost more than a plain input token. Sol writes are $6.25, Terra $2.50, Luna $0.25. That's 1.25x the standard input price, so priming a cache costs you a 25% premium on the first pass in exchange for 90% off every read after it.
The July 30, 2026 changelog entry matters because the two cuts weren't the same size: "Starting July 30, GPT-5.6 Luna costs 80% less, while GPT-5.6 Terra costs 20% less." Luna got the headline. If you sized a budget off a Terra quote from July, you're only 20% off. If you sized one off Luna, your estimate is five times too high, which is the nicer direction to be wrong in.
All three share the same envelope: 1.05M token context window, 128K max output, knowledge cutoff February 16, 2026. So any of them fits your documents — but fitting and costing the same are different things. The 272K band boundary above applies to all three equally, so the choice between models is about cost and quality, while the choice about how much context you send is a separate pricing decision of its own.
Legacy models are still billed at their own rates
This matters more than it sounds. When a vendor retires a model slug, one of two things happens: the request errors, or it silently redirects and bills at the successor's rate. I checked every older row on OpenAI's table against its GPT-5.6 replacement and none of them have been quietly repriced, even the ones already carrying a shutdown date. You pay what the row says.
| Model | Input | Cached input | Output |
|---|---|---|---|
| gpt-5 | $1.25 | $0.125 | $10.00 |
| gpt-5-mini | $0.25 | $0.025 | $2.00 |
| gpt-5-nano | $0.05 | $0.005 | $0.40 |
| gpt-5-pro | $15.00 | — | $120.00 |
| gpt-4.1 | $2.00 | $0.50 | $8.00 |
| gpt-4.1-mini | $0.40 | $0.10 | $1.60 |
| gpt-4.1-nano | $0.10 | $0.025 | $0.40 |
| gpt-4o | $2.50 | $1.25 | $10.00 |
| gpt-4o-mini | $0.15 | $0.075 | $0.60 |
| gpt-3.5-turbo | $0.50 | — | $1.50 |
| o1 | $15.00 | $7.50 | $60.00 |
| o1-pro | $150.00 | — | $600.00 |
| o3 | $2.00 | $0.50 | $8.00 |
| o3-pro | $20.00 | — | $80.00 |
| o3-mini | $1.10 | $0.55 | $4.40 |
| o4-mini | $1.10 | $0.275 | $4.40 |
o1-pro at $150 input and $600 output per million is still sitting there, roughly thirty times Sol's input price for a model that shuts down in October. Nobody should be sending it traffic. Someone is.
Two other things fall out of this table. gpt-4o's cached input is $1.25, only half its standard input rate rather than the 90% discount the GPT-5.6 rows get, so caching is worth dramatically less on the gpt-4o pair specifically — the gpt-5, gpt-5-mini and gpt-5-nano rows still show the full 90% ratio. And the legacy rows show a dash in the cache-writes column, meaning the write premium is a GPT-5.6-era addition rather than something you were already paying.
Fast mode is Priority Processing with a new name
Do not treat this as a new feature to evaluate. OpenAI's own guide says: "Priority processing was renamed Fast mode on July 30, 2026." Same tier, new label, generally available right now. The old service_tier: "priority" value still works and is an alias for service_tier: "fast", so nothing in your code breaks.
What did change is speed. OpenAI says it "increased the speed at which Fast mode operates for gpt-5.6-sol to make it up to 2.5× faster than Standard processing," with "more consistent latency while keeping pay-as-you-go flexibility."
The price is the interesting part. For the entire GPT-5.6 family, Fast mode is exactly 2x Standard, to the cent — and that holds in both context bands. Short-context Fast prices:
| Model | Fast input | Fast cached | Fast output |
|---|---|---|---|
| gpt-5.6-sol | $10.00 | $1.00 | $60.00 |
| gpt-5.6-terra | $4.00 | $0.40 | $24.00 |
| gpt-5.6-luna | $0.40 | $0.04 | $2.40 |
Above 272K input tokens the same 2x applies to the long-context rates: Sol runs $20.00 in / $90.00 out, Terra $8.00/$36.00, Luna $0.80/$3.60.
Older models get a smaller markup. gpt-4o goes from $2.50 to $4.25 input, which is 1.7x. o3 goes $2.00 to $3.50, or 1.75x. gpt-4o-mini goes $0.15 to $0.25, about 1.67x. If your latency-sensitive path is still on gpt-4o, Fast mode is proportionally cheaper there than it would be on the new family — the sort of asymmetry that quietly changes the shape of a migration plan. (None of the older models publishes a long-context price in any tier — gpt-5.5, gpt-5.5-pro, gpt-5.4 and gpt-5.4-pro are explicitly labelled "(<272K context length)", the rest simply carry no context band at all — so for them the Fast price you see is the whole story.)
Fast mode isn't available on everything. The pricing table's fast tier lists gpt-5.6-sol, terra and luna, then gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.2, gpt-5.1, gpt-5, gpt-5-mini, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-4o, gpt-4o-2024-05-13, gpt-4o-mini, o3 and o4-mini. No other slugs appear with populated prices.
Batch and Flex: the same 50%, two different trades
Batch is 50% off with a 24-hour completion window, plus what OpenAI calls "a separate pool of significantly higher rate limits." For Sol that's $2.50 input, $0.25 cached, $3.125 cache write, $15.00 output. Terra lands at $1.00/$0.10/$1.25/$6.00 and Luna at $0.10/$0.01/$0.125/$0.60. The long-context band halves the same way — Sol at $5.00 in / $22.50 out above 272K input tokens. Exact halves in both bands, no rounding surprises.
Flex charges the same halved rates for the three GPT-5.6 models — I checked those rows cell by cell, in both bands — but the two tiers are not mirror images: Batch prices 34 model rows where Flex prices 16, so eighteen slugs have a Batch price and no Flex option at all. (The two easy to miss on the Flex list are gpt-5.5-pro and gpt-5.4-pro, both at $15.00 in / $90.00 out — half their standard $30/$180.) And Flex carries a caveat Batch doesn't: OpenAI's guide says plainly that "Flex processing is in beta with limited model availability." What you're buying is different too. Batch trades turnaround for money; Flex trades per-request latency for money and warns about "occasional resource unavailability." So Flex is for synchronous work where you can tolerate slow and can retry, and Batch is for work you can hand over and collect tomorrow. Same bill, very different failure mode.
Prompt caching, and a change that will bite migrations
Caching turns on by itself for any prompt of 1024 tokens or longer. No flag, no opt-in. On the GPT-5.6 family that's a 90% discount on the cached portion, which is the single biggest lever on this whole page if you run long system prompts.
The gotcha is a behaviour change, not a price. OpenAI's caching guide states that GPT-5.6 and later "cache exact prompt prefixes at cache breakpoints," with an implicit breakpoint at the latest user or tool message, and that unlike earlier models it "does not automatically fall back to the longest matching unmarked prefix before that breakpoint." Translated into a bill: a prompt-assembly pattern that was reliably hitting cache on GPT-5 can quietly stop hitting it on Sol, and you find out from your invoice rather than from an error. Watch your cached-token counts for the first week after you switch.
One more line item that's easy to miss: regional processing endpoints carry a 10% uplift for models released on or after March 5, 2026 that are eligible for data residency. Every price above assumes you're not using them.
Usage tiers
Access is gated by cumulative spend, and promotion is automatic — the docs say "as your spend on our API goes up, we automatically graduate you to the next usage tier."
| Tier | Qualification | Monthly usage limit |
|---|---|---|
| Free | Allowed geography | $100 / month |
| Tier 1 | $5 paid | $100 / month |
| Tier 2 | $50 paid | $500 / month |
| Tier 3 | $100 paid | $1,000 / month |
| Tier 4 | $250 paid | $5,000 / month |
| Tier 5 | $1,000 paid | $200,000 / month |
Spend is the only variable on that page. There's no waiting period listed, no account-age criterion, nothing about payment history.
Two things I went looking for and could not find. First, per-model RPM and TPM numbers: I checked the rate-limits guide, the models page and the pricing page on 10 August, and none of them carry a public per-tier RPM/TPM table. The rate-limits guide points you at the models page for "a high-level summary," but the models page's public HTML has no such figures, and the authoritative per-org numbers live behind login at platform.openai.com/settings/organization/limits, which returns 403 without an account. Second, a new-account free credit: I checked the rate-limits guide, the quickstart and the pricing page, and none of them state a dollar grant for new signups. The widely repeated "$5 free credit" isn't on any of those pages, so I'm not printing it as a number. The only "Free" is that top table row, and it's a ceiling, not a gift.
Dated shutdowns you need in your calendar
These are absolute dates from OpenAI's deprecations page. I'm writing them out in full because a page that says "next month" is worthless six weeks later.
| Date | What goes away | Replacement |
|---|---|---|
| 10 August 2026 | gpt-5.2-chat-latest, gpt-5.3-chat-latest | gpt-5.6-sol |
| 26 August 2026 | Assistants API | Responses API + Conversations API |
| 24 September 2026 | Videos API, sora-2, sora-2-pro and all their snapshots | none listed |
| 28 September 2026 | gpt-3.5-turbo-instruct, gpt-3.5-turbo-1106, babbage-002, davinci-002 | gpt-5.6-terra |
| 23 October 2026 | gpt-3.5-turbo (base and 0125), gpt-4 and gpt-4-turbo families, gpt-4.1-nano, gpt-4o-2024-05-13, gpt-image-1, o1, o1-pro, o3-mini, o4-mini | GPT-5.6 family; gpt-image-1 → gpt-image-2 |
| 30 November 2026 | v1/prompts reusable prompt objects, the Evals dashboard and API (read-only from 31 October), Agent Builder | app code / Agents SDK |
| 1 December 2026 | gpt-image-1.5, gpt-image-1-mini, chatgpt-image-latest | gpt-image-2 |
| 11 December 2026 | gpt-5-2025-08-07, gpt-5-mini-2025-08-07, gpt-5-nano-2025-08-07, gpt-5-pro-2025-10-06, o3-2025-04-16, o3-pro-2025-06-10 | Sol / Terra / Luna |
| 20 January 2027 | gpt-realtime, gpt-audio, gpt-4o-audio, gpt-4o-realtime and the mini variants | gpt-realtime-2.1 (mini variants → gpt-realtime-2.1-mini), gpt-audio-1.5; gpt-4o-mini-transcribe → its 2025-12-15 snapshot |
Mind the gpt-3.5 split: the -instruct and -1106 variants go on 28 September, a month before the rest of the family. If you're on either — both are still individually priced today — the October date everyone quotes is a month late for you.
The two chat-latest snapshots went dark on 10 August 2026, the same day this article was drafted. Notice also that the December wave maps gpt-5-pro and o3-pro onto gpt-5.6-sol with reasoning.mode: pro rather than onto a separate pro model.
Fine-tuning is winding down too. The pricing page says the platform "is no longer accessible to new users," and the deprecations page puts a hard date on the rest: active existing customers can no longer create new fine-tuning jobs after 6 January 2027. Fine-tuned models stay available for inference until their base models are deprecated. If you're already on it, o4-mini fine-tuning is $100.00/hour to train with inference at $4.00/$1.00/$16.00 per 1M, halved to $2.00/$0.50/$8.00 if you opt into data sharing. That data-sharing discount is the only one of its kind on the page, and it applies to fine-tuned inference only, not to regular chat calls.
Everything else on the bill
Embeddings are cheap enough to ignore in most budgets: text-embedding-3-small is $0.02 per 1M, 3-large is $0.13, and ada-002 is $0.10 — which means the old ada model now costs five times the current small one. Moderation via omni-moderation-latest is free, listed literally as "Free" in the input column.
Image generation bills tokens in two streams — and both current models have a shutdown date. gpt-image-1.5 charges $8.00 input, $2.00 cached and $32.00 output per 1M image tokens, plus $5.00/$1.25/$10.00 on the text side. gpt-image-1-mini is $2.50/$0.25/$8.00 for image tokens and $2.00/$0.20 for text input and cached input. Both are removed from the API on 1 December 2026 along with chatgpt-image-latest, replaced by gpt-image-2 — which is already priced on the page at $8.00/$2.00/$30.00 per 1M image tokens ($5.00/$1.25 text). If you're building on image generation in the autumn, build on gpt-image-2 from the start.
Last housekeeping note, and it trips up scripts more than people: platform.openai.com/docs/* now 301-redirects to developers.openai.com/api/docs/*. Old bookmarks resolve fine, so nothing looks broken, but if you've got a scraper or an agent pinned to the old host it's following a redirect it may not be logging. Same class of quiet migration as the retired Grok slugs that keep billing without erroring.
If you're comparing this against what other vendors charge, the free AI API rate limits page covers the zero-dollar end of the market, and whether the Claude API is free does the same for Anthropic.
Top comments (0)